Geospatial Analysis of High-Crash Intersections and Rural Roads using Naturalistic Driving Data

Transcription

1 11-UT-013 Geospatial Analysis of High-Crash Intersections and Rural Roads using Naturalistic Driving Data Final Report Brad R. Cannon Jeremy Sudweeks Submitted: September 26, 2011 i

2 ACKNOWLEDGMENTS The authors of this report would like to acknowledge the support of the stakeholders of the National Surface Transportation Safety Center for Excellence (NSTSCE): Tom Dingus from the Virginia Tech Transportation Institute, John Capp from General Motors Corporation, Carl Andersen from the Federal Highway Administration, Chris Hayes from Travelers Insurance, Martin Walker from the Federal Motor Carrier Safety Administration, and Jose Gomez from the Virginia Department of Transportation and the Virginia Center for Transportation Innovation and Research. The NSTSCE stakeholders have jointly funded this research for the purpose of developing and disseminating advanced transportation safety techniques and innovations. The Government and others acting on its behalf shall have unlimited rights to obtain, reproduce, publish or otherwise use the data developed in the performance of this cooperative agreement pursuant to 49 CFR Part

3 EXECUTIVE SUMMARY Despite the fact that overall road safety continues to improve, intersections and rural roads persist as trouble areas or hotspots. Using a previously developed method, naturalistic driving data were identified through intersection and rural road hotspots and compared to naturalistic driving data through similar intersections and rural road locations, but with low crash counts. Few significant differences were found between driver behaviors in the low-crash and high-crash areas of study. For the few significant differences, there was not an apparent consistent pattern. A compelling possibility that warrants further investigation is that drivers may not recognize when they have entered an area with high-crash counts and that a change in driving behavior would be appropriate. i

4

5 TABLE OF CONTENTS EXECUTIVE SUMMARY... i LIST OF FIGURES... v LIST OF TABLES... vii LIST OF ABBREVIATIONS... ix CHAPTER 1. INTRODUCTION... 1 CHAPTER 2. METHODS... 3 IDENTIFYING RURAL ROAD HOTSPOTS... 3 Selecting Hotspots with Naturalistic Driving Data for Analysis... 4 IDENTIFYING INTERSECTION HOTSPOTS Intersection Selection VIDEO REDUCTION Reductionist Training Event Validation Question Reduction Spot-Checking ANALYSIS METHODS CHAPTER 3. RESULTS DRIVER BEHAVIORS Secondary Tasks SUMMARY OF ANALYSIS CHAPTER 4. DISCUSSION AND CONCLUSIONS FUTURE WORK LIMITATIONS OF THE STUDY APPENDIX A REFERENCES iii

6

7 LIST OF FIGURES Figure 1. Diagram. Flow chart of high-crash road segment selection process Figure 2. Aerial Image. Prices Fork Road contiguous high-crash (red) and low-crash (blue) road segments. (Source: )... 6 Figure 3. Aerial Images. US-460 separated high-crash (red) and low-crash (blue) road segments. (Source: )... 7 Figure 4. Aerial Images. VA-28 separated high-crash (red) and low-crash (blue) road segments. (Source: )... 8 Figure 5. Aerial Image. VA-7 Opposing Direction High-Crash (Red) and Low-Crash (Blue) Road Segments. (Source: )... 9 Figure 6. Chart. Crash type counts per road segment Figure 7. Aerial Image. Braddock Rd at Backlick Rd (node ). This intersection served as a high-crash intersection in the analysis. (Source: ).12 Figure 8. Aerial Image. Lee Hwy at Waples Mill Rd and Shirley Gate Rd (node ). This intersection served as a high-crash intersection in the analysis. (Source: )...13 Figure 9. Aerial Image and Map. Waxpool Rd. at Farmwell Rd (node ). This intersection served as a low-crash intersection in the analysis. (Source: )...14 Figure 10. Aerial Image. Balls Ford Rd at Sudley Rd. This intersection served as a lowcrash intersection during analysis. (Source: )...15 Figure 11. Chart. Crash types reported at two intersections between the years 1999 and Figure 12. Chart. Crash types reported at two intersections between the years 1999 and Figure 13. Chart. Major crash factors as reported in VDOT crash tables v

8

9 LIST OF TABLES Table 1. Summary measures used in selection of high-crash intersections. Gray shading indicates high-crash intersections used in analysis Table 2. Summary measures used in selection of low-crash intersections Table 3. Number of trips analyzed per roadway and segment type Table Car aggregate 95% confidence intervals Table 5. Older Driver aggregate 95% confidence intervals Table 6. VA-7 95% confidence intervals Table 7. VA-28 95% confidence intervals Table 8. US % confidence intervals Table 9. Prices Fork Road 95% confidence intervals vii

10

11 LIST OF ABBREVIATIONS AADT ESRI GIS GPS SQL UA UC VDOT VTTI Average Annual Daily Traffic Environmental Systems Research Institute, Inc. Geographic Information System Global Positioning System Structured Query Language Urbanized Area Urban Cluster Virginia Department of Transportation Virginia Tech Transportation Institute ix

12

13 CHAPTER 1. INTRODUCTION As overall road safety continues to improve, some problem areas continue to have a disproportionate number of fatal and injurious crashes. Two types of areas that often have an unusually high number of injuries and fatalities are intersections and rural areas. Recently, naturalistic driving data have been able to provide a new opportunity for exploring crashes in more detail than was previously possible with only crash report data. Naturalistic data also provide the possibility of studying drivers behavior during non-crash situations. This allows comparisons of behavior while driving through an area in both crash and non-crash situations and comparisons of driving behavior in different areas. Geographic information systems (GIS) allow analysis based on geographical location variables (i.e., latitude and longitude from Global Positioning System [GPS]) contained within the data. Data containing geographical information are known as geospatial data. Previous work developed a method using GIS to find rural driving within large naturalistic driving datasets. (1) In the latest work, naturalistic data through high-crash rural roads were identified and compared to naturalistic data through low-crash rural roads. This method was also adapted and used to compare naturalistic data through high-crash intersections with naturalistic data from low-crash intersections. 1

14

15 CHAPTER 2. METHODS In previous work, a method was developed to identify naturalistic driving data in rural areas. (1) That method was used here to identify rural driving in the two naturalistic driving datasets used in this work. These processes were also adapted to identify driving through intersections in the same datasets. The naturalistic driving data used were collected in the 100-Car Study (2) and the Virginia Tech Transportation Institute s (VTTI) Older Driver Collection. The 100-Car Study data were collected in the northern Virginia/Washington, DC metro area and provide a resource of naturalistic driving data for a variety of drivers in different situations. Due to the collection area s urban location, most of the driving data in the dataset are on urban roads, but still provide some driving data for rural areas. The Older Driver data were collected from drivers in and around Blacksburg in the New River Valley area of southwestern Virginia. The Older Driver data were used to augment the 100-Car data because of a higher prevalence of rural driving in southwestern Virginia than in northern Virginia. The method for identifying rural driving is dependent on the use of GIS software. In the naturalistic driving data, the latitude and longitude pair that identifies the current location of the car is recorded at a rate of 1 Hz. Collectively, these points are referred to as breadcrumb trails. The definition of a rural road used here is the same as the definition used in the earlier study (1), which is a road outside of an urbanized area (UA) or urban cluster (UC) as defined by the U.S. Census Bureau, excluding interstate highways. Consequently, breadcrumb trails outside the boundaries of any UA or UC and not falling on an interstate highway were identified for study in this work. Geographic data describing the boundaries of UAs and UCs were obtained from the Census Bureau in the shapefile format, which stores geographic data for use with commonly available GIS software. Shapefiles were also acquired from the Virginia Department of Transportation (VDOT) which describe the centerlines of the roadways in Virginia. Tables describing crashes in Virginia from 1999 through 2005 were also obtained from VDOT. These tables included details such as the latitude and longitude of the crash, the type of crash, and the number of vehicles involved. IDENTIFYING RURAL ROAD HOTSPOTS Choosing the problematic rural locations for specific study required the development of a method for identifying problem areas or hotspots. The first attempt at doing this made use of the point density tool provided in the Environmental Systems Research Institute s (ESRI) ArcGIS Desktop software. The latitude and longitude from the VDOT crash tables were plotted in ArcGIS and the densities of these crash locations were calculated using the point density tool. The initial results of using this tool made it difficult to identify rural hotspots because the numbers of crashes in urban areas outnumbered the crashes in rural areas to such a large degree that the differences in crash counts between individual rural areas was indistinguishable. This process was refined by excluding the crash points within UAs and UCs or on interstate highways. In this way, only the crashes that are on rural roads using the rural-road definition described previously were used to calculate the crash densities of areas around the state. While this method could show promise for identifying a jurisdiction with a crash density problem, it 3

16 identifies an area where crashes tend to cluster, and this could include multiple roads as well as multiple road segments. For example, the method frequently identified areas where roads intersected, particularly when several other roads were in the near vicinity. This result may not clearly indicate whether or not a certain approach to the intersection, or a certain road, could be considered the hotspot. It could also identify an area as a hotspot simply because many roads are close together, even though the individual road segments have typical crash counts. To more specifically identify a road segment with a crash problem, a different method was needed. A spatial join was used to associate each crash with the nearest road segment. This was done by joining the crashes in the VDOT crash data with the nearest road segment in VDOT s centerline road data. The result of the spatial join was a geospatial table in a PostgreSQL database. By writing the result to a database table, a count of the crashes and crash types on each road segment could be obtained through simple SQL queries. Selecting Hotspots with Naturalistic Driving Data for Analysis The 100 road segments with the most crashes were mapped and overlaid with 100-Car and Older Driver breadcrumb data to find segments with large enough numbers of naturalistic driving traversals to provide a sample size sufficient for analysis. A spatial join was used to associate the breadcrumbs from the 100-Car and Older Driver datasets with the nearest (within 200 ft) road segment with the top 100 crash counts. This result was also written to a geospatial table in a PostgreSQL database. Using SQL queries, the number of traversals through each road segment by each driver was used to identify the hotspots for further review by data reductionists. The selection of the high-crash road segments process is shown in Figure 1. Another SQL query was used to randomly choose trips through the chosen high-crash road segments. The identifying trip File ID and frame numbers were given to the data reductionists in a database table for question reduction as discussed below. 4

17 VDOT Crash Locations Spatial Join Crashes in Rural Areas Census Bureau Urbanized Areas and Urban Clusters VDOT Road Centerline Data Spatial Join Crashes and Associated Road Segments Aggregate SQL Query 100-Car and Older Driver Naturalistic Driving Data Breadcrumbs Count of Crashes on Each Rural Road Segment Rank Segments by Crash Count Top 100 Crash Rural Road Segments Spatial Join and Aggregate SQL Query Count of Trips on Top 100 Crash Rural Road Segments Choose Segments with Multiple Trips by Multiple Drivers for Data Reduction Figure 1. Diagram. Flow chart of high-crash road segment selection process. Two high-crash rural roadway segments in northern Virginia and two additional high-crash rural roadway segments in southwestern Virginia were chosen for analysis. All four segments had crash counts in the top 100 of rural roadway segments in the state and had similar lengths. For comparison, each of the four high-crash segments was matched up with a similar low-crash segment. To control for traffic volumes, nearby segments on the same roadway were chosen. This also helped find low-crash road segments with geometries similar to their high-crash counterparts. In one case, the low-crash and the high-crash segments were contiguous segments. In two cases, they were separated but near each other. In the final case, the high-crash segment was the eastbound lanes of a divided roadway. The westbound lanes were on a low-crash segment and were chosen for comparison to the high-crash eastbound lanes. The road segments used in the analysis were on Prices Fork Road in Montgomery County (Figure 2), US-460 in Giles County (Figure 3), VA-28 in Prince William County (Figure 4), and VA-7 in Loudoun County (Figure 5). 5

18 Figure 2. Aerial Image. Prices Fork Road contiguous high-crash (red) and low-crash (blue) road segments. (Source: ) The high-crash segment of Prices Fork Road is a two-lane road with no shoulder and several horizontal and vertical curves. The segment has center and side lane markings. There are a number of side roads and driveways (gravel and paved) that branch from the road segment. The road segment is generally surrounded by trees and steep hillsides with a few homes and small businesses. The low-crash segment is very similar to the high-crash segment, but with more straight driving and not as many steep hillsides. The high-crash and low-crash segments are contiguous segments and are shown in the aerial image in Figure 2. 6

19 Figure 3. Aerial Images. US-460 separated high-crash (red) and low-crash (blue) road segments. (Source: ) The high-crash segment of US-460 is a section of the westbound lanes where US-460 is a divided highway with two lanes traveling in each direction. There are several un-signalized intersections throughout the road segment. There is a solid lane marking on the right side of the right lane and the left side of the left lane with a dashed line dividing the two lanes. There is a shoulder to the right of the right lane. Some parts of the road segment are surrounded by trees and hills, while other parts have an open view and homes or small businesses. The low-crash segment of US-460 is similar to the high-crash segment. The low-crash segment has more open views and fewer intersections. The traffic on the low-crash segment is also traveling in a primarily westbound direction away from the Blacksburg UA. The high-crash and low-crash segments are separated segments, but are near each other. They are shown in the aerial images in Figure 3. 7

20 Figure 4. Aerial Images. VA-28 separated high-crash (red) and low-crash (blue) road segments. (Source: ) The high-crash segment of VA-28 is a two-lane road with a single solid line on each edge. The road is partially divided by double yellow lines, dashed lines, mixed dash and solid, and mixed solid and dash. The roadway is surrounded at times by trees but includes some sections with open views. The road segment is mostly straight driving with one traffic signal toward the end and some un-signaled intersections throughout the segment. There are some homes and driveways off of the road. The low-crash segment is very similar to the high-crash segment with more trees surrounding the roadway and a shoulder for most of the roadway. The high-crash and low-crash segments are separated, but are near each other. They are shown in the aerial images in Figure 4. 8

21 Figure 5. Aerial Image. VA-7 Opposing Direction High-Crash (Red) and Low-Crash (Blue) Road Segments. (Source: ) The high-crash segment of VA-7, shown in red in Figure 5, is one side of a divided highway with two lanes traveling in the eastbound direction. There is an on-ramp at the beginning and at the end of the segment. There is a solid line on the left side of the left lane and one on the right side of the right lane. The two lanes are separated by a dashed line. There is a shoulder to the right of the right lane. The road segment is mostly straight driving with some wide turns and short trees along the road, but has an open view for the most part. The low-risk road segment parallels the high-crash segment, containing the westbound lanes. This segment is almost identical to the high-crash segment. Note that the high-crash and low-crash segments are both on a bypass section of VA-7. A portion of a business loop of VA-7 can also be seen in Figure 5 splitting from the bypass road segments. A count of the crash types was calculated for each of the high-crash and low-crash road segments by using an aggregate SQL query on the geospatial table created from the spatial join of the VDOT crash tables and the VDOT centerline data. The graph of this result is shown in Figure 6. The terms in Figure 6 are from VDOT s crash tables. A non-collision crash is a crash where the vehicle runs off the road and does not collide with another object. 9

22 Figure 6. Chart. Crash type counts per road segment. IDENTIFYING INTERSECTION HOTSPOTS The process of retrieving intersection traversals for evaluation in this study followed a two-stage process in which the first stage was to identify intersections of interest, and the second stage was to select individual instances of drivers in the naturalistic data passing through the selected intersections. Intersection Selection The selection of intersections of interest was based on the number and types of crashes reported by VDOT and the coverage of the 100-Car dataset at these intersections. The 100-Car data were used in the intersection analysis because more of the intersection hotspots were in northern Virginia where the 100-Car data are more extensive. VDOT provided a listing of crashes for the years 2005 through 2007 at approximately 6,000 intersections, along with a ranking for each of the intersections based on the total number of fatal or injury crashes over this three-year period. ArcMap was used to create 200-ft buffers around each intersection s latitude-longitude pairing and to identify which of the 100-Car trips passed through the intersection. Summary measures 10

23 were calculated for each intersection to estimate how many usable 100-Car trips passed through the intersection, how many distinct drivers passed through the intersection, and the largest contribution that a single driver made toward the total number of traversals. High-Crash Intersections Focus for selecting high-crash intersections was restricted to the 15 intersections with the largest number of fatal or injury crashes in the 2005 through 2007 timeframe. Table 1 below provides a listing of the summary measures relevant to high-crash intersection selection. Table 1. Summary measures used in selection of high-crash intersections. Gray shading indicates high-crash intersections used in analysis. Fatal / Injury Rank (frequency) Node Number Fatal / Injury Crash Count Number 100-Car Trips Number Distinct Drivers Percentage of Largest Participant Contribution Fatal / Injury Percentage of Total Crashes It was determined that a single participant should not contribute more than 30% of the total observed intersection crossings in the 100-Car data and that fatal and injury crashes should account for at least 60% of the total crashes reported at the intersection. Based upon these criteria, the large number of 100-Car trips and relatively large number of distinct drivers observed, the following intersections were selected: Braddock Rd at Backlick Rd (node ) and Lee Hwy at Waples Mill Rd and Shirley Gate Rd (277764). Figure 7 and Figure 8 provide aerial imagery of the selected intersections. 11

24 Figure 7. Aerial Image. Braddock Rd at Backlick Rd (node ). This intersection served as a high-crash intersection in the analysis. (Source: ) 12

25 Figure 8. Aerial Image. Lee Hwy at Waples Mill Rd and Shirley Gate Rd (node ). This intersection served as a high-crash intersection in the analysis. (Source: ) Low-Crash Intersections In order to provide a normative comparison for the high-crash intersections, low-crash intersections were identified. Although traffic density measures such as average annual daily traffic (AADT) were not explicitly used in the normative intersection selection, efforts were made to select intersections comparable to the high-crash intersections. The selection criteria for low-crash intersections included the following: total fatal or injury crashes at low-crash candidate intersections do not exceed 50% of the high-crash intersection total count; similar geometry and curvature on approach legs; a single participant does not contribute more than 30% of the total observed intersection crossings in the naturalistic driving data; a large number of 100-Car trips; and a relatively large number of distinct drivers. Based on these criteria, the following intersections were selected as low-crash intersections: Waxpool Rd at Farmwell Rd (713920) and Balls Ford Rd at Sudley Rd (546133). Table 2 contains the summary measures used in selecting a low-crash intersection. 13

26 Table 2. Summary measures used in selection of low-crash intersections. Fatal / Injury Rank (frequency) Node Number Fatal / Injury Crash Count Number of 100- Car Trips Number of Distinct Drivers Percentage of Largest Participant Contribution Fatal / Injury Percentage of Total Crashes It can been seen in Table 1 and Table 2 that, although the low-crash intersections have a number of fatal or injury crashes, these numbers are significantly less than those for the high-crash intersections. Figure 9 and Figure 10 provide aerial imagery of the selected low-crash intersections. Figure 9. Aerial Image and Map. Waxpool Rd. at Farmwell Rd (node ). This intersection served as a low-crash intersection in the analysis. (Source: ) 14

27 Figure 10. Aerial Image. Balls Ford Rd at Sudley Rd. This intersection served as a lowcrash intersection during analysis. (Source: ) The graphs in Figure 11 and Figure 12 provide a breakdown of the crash types reported at each location for the years 1999 through Earlier (Figure 6), raw counts of crashes had been used to identify road segments with high crashes. Percentages were used here to compare crash types between intersections. The high-crash intersections had been previously identified by VDOT using raw counts. 15

28 Figure 11. Chart. Crash types reported at two intersections between the years 1999 and

29 Figure 12. Chart. Crash types reported at two intersections between the years 1999 and Epoch Selection The process of selecting individual intersection traversals for inclusion in the analysis was straightforward and simple. Based on previous reduction efforts, files with undesirable attributes (such as obstructed video views) were excluded from consideration. A random sample of 120 files per intersection node were drawn from the remaining files and presented to video analysts for video validation and reduction. VIDEO REDUCTION Following the identification of 100-Car files with traversals through roadway segments or intersections of interest, video analysis was completed. This video analysis, also referred to as video reduction, was completed in order to collect information about the driving environment, driver states (e.g., inattentive or impaired), driver behavior, and roadway infrastructure elements. Trained video analysts completed the video reduction in secured computing labs using VTTI 17

30 proprietary software. The following sections briefly describe the different types of video analysis completed, protocol development, and analyst training. Reductionist Training As the video reductionists employed in this task were quite experienced, training needs were minimal. Training consisted primarily of familiarizing reductionists with the objectives of the research task, discussing differences in video fidelity between the datasets, and reviewing the reduction protocol. Although certain questions were only applicable to roadway segments or intersections, there was one protocol provided for the video reduction. Aspects of this reduction task that likely differed from previous research efforts were emphasized during the review of the reduction protocol. After meeting with a researcher for the brief training session, reductionists were instructed to read and familiarize themselves with the reduction protocol prior to beginning reduction. Questions encountered during reduction were addressed to a researcher or the reduction manager. Event Validation The first step in the reduction process was to determine if a trip that had been labeled as passing along a roadway or intersection was, in fact, of interest. As can occasionally happen with naturalistic data collection, there are times at which the video quality does not allow for reliable video analysis. In addition, the GPS measures collected by the instrumentation package are at times not amenable to accurate geospatial processing, resulting in trips that seem too short or long for a particular road segment. In such cases, the reductionist does not complete the video analysis and the trip is excluded from further analysis. As is mentioned in the section describing trip selection, additional trips are sampled to account for the possibility of trips being discarded due to unusable data. Question Reduction The question reduction was a detailed analysis that sought to address several factors relevant to the traversal. Questions in the reduction were grouped according to the factor being addressed; for example: video quality, driving environment, driver states (e.g., inattentive or impaired), driver behavior, and roadway infrastructure. To the extent possible, questions included in this reduction were based upon reductions completed in previous studies. This provided the opportunity to make use of questions that have been successfully used in previous research efforts and with which the reductionists are familiar. The question reduction protocol is listed in the appendix. Spot-Checking The primary mechanism used to encourage consistent responses across reductionists was frequent spot-checking. Each reductionist would spend a small amount of time during each shift reviewing the work of another reductionist. Discrepancies between the reductionists were noted and the reductionists would discuss reduction differences in order to converge on consistent use of reduction values. 18

31 ANALYSIS METHODS The focus of the analyses conducted in this study was on simple comparisons between high- and low-crash road segments across a number of measures. These comparisons were conducted at two different levels of aggregation, the first of which disregarded the specific roadway but simply made comparisons between high-crash and low-crash segments. For example, the VA-7 and VA-28 segments were simply collapsed into high- and low-crash groupings. The second comparison matched each segment within the roadway from which it was selected. For example, the VA-7 high-crash segment was compared only to the VA-7 low-crash segment. Table 3 displays the number of trips used in the analysis for each of the levels of aggregation. Table 3. Number of trips analyzed per roadway and segment type. Roadway Naturalistic Dataset Segment Type Number of Trips VA Car High-crash 81 VA Car Low-crash 94 VA Car High-crash 56 VA Car Low-crash 89 US-460 Older Driver High-crash 48 US-460 Older Driver Low-crash 38 Prices Fork Rd Older Driver High-crash 61 Prices Fork Rd Older Driver Low-crash 95 Braddock & Backlick 100-Car High-crash 91 Lee Hwy & Waples Mill 100-Car High-crash 107 Waxpool & Farmwell 100-Car Low-crash 60 Balls Ford & Sudley 100-Car Low-crash 102 As can be seen in Table 3, although sampling efforts sought to have at least 100 trips for each roadway and segment type, this could not be met in all cases. In particular, US-460 has a low number of trips possible, resulting in an unpowered analysis. The measures used in analysis were grouped into those that address roadway infrastructure, and those that address driver states and behaviors. Driver state and behavior variables that were analyzed include the following: secondary task engagement; driver impairment; unsafe, illegal, or reckless actions; hand placement on the steering wheel; and seatbelt use. The roadway infrastructure variables considered include the following: roadway elements such as sight distance or roadway delineation; visual obstructions; traffic density; and lighting conditions. As can be seen in Figure 13, these categories align well with the most frequent major crash factors reported in the VDOT crash tables. 19

32 Figure 13. Chart. Major crash factors as reported in VDOT crash tables. 20

33 CHAPTER 3. RESULTS DRIVER BEHAVIORS This section describes the frequency of various driver behaviors, such as secondary task engagement, observed in each of the high- and low-crash road segments. Findings are reported separately for the 100-Car and Older Driver datasets. Secondary Tasks Table 4 through Table 9 provide comparisons of the crash segments using distraction-related measures. The comparisons include measures related to number and duration of tasks, as well as types of tasks. Table 4 focuses on the 100-Car dataset while Table 5 focuses on older drivers. Table 6 through Table 9 provide distraction comparisons within individual road segments. Distraction Type Table Car aggregate 95% confidence intervals. Low-crash Mean or Proportion High-crash Mean or Proportion Difference Lower Bound Upper Bound Segment with Higher Distraction Involvement Number of secondary tasks per traversal Low-crash Length of longest duration secondary task (sec) Number of distinct secondary tasks per traversal passenger distraction cognitive distraction Not significant Not significant Not significant Not significant object in vehicle distraction Low-crash electronic device distraction High-crash vehicle controls distraction Not significant external distraction Low-crash eating distraction Not significant other distraction High-crash 21

34 Distraction Type Table 5. Older Driver aggregate 95% confidence intervals. Low-crash Mean or Proportion High-crash Mean or Proportion Difference Lower Bound Upper Bound Segment with Higher Distraction Involvement Number of secondary tasks per traversal Not significant Length of longest duration secondary task (sec) Not significant Number of distinct secondary tasks per traversal Not significant passenger distraction Not significant cognitive distraction Not significant object in vehicle distraction Not significant electronic device distraction NA NA NA NA NA Not significant vehicle controls distraction Not significant external distraction Low-crash eating distraction Not significant other distraction NA NA NA NA High-crash 22

35 Distraction Type Table 6. VA-7 95% confidence intervals. Low-crash Mean or Proportion High-crash Mean or Proportion Difference Lower Bound Upper Bound Segment with Higher Distraction Involvement Number of secondary tasks per traversal Not significant Length of longest duration secondary task (sec) Not significant Number of distinct secondary tasks per traversal Not significant passenger distraction Not significant cognitive distraction Not significant object in vehicle distraction Not significant electronic device distraction Not significant vehicle controls distraction Not significant external distraction Low-crash eating distraction Not significant other distraction Not significant 23

36 Distraction Type Table 7. VA-28 95% confidence intervals. Low-crash Mean or Proportion High-crash Mean or Proportion Difference Lower Bound Upper Bound Segment with Higher Distraction Involvement Number of secondary tasks per traversal Low-crash Length of longest duration secondary task (sec) Not significant Number of distinct secondary tasks per traversal Not significant passenger distraction Not significant cognitive distraction Not significant object in vehicle distraction Low-crash electronic device distraction High-crash vehicle controls distraction Not significant external distraction Low-crash eating distraction Not significant other distraction NA NA NA NA High-crash 24

37 Distraction Type Table 8. US % confidence intervals. Low-crash Mean or Proportion High-crash Mean or Proportion Difference Lower Bound Upper Bound Segment with Higher Distraction Involvement Number of secondary tasks per traversal Not significant Length of longest duration secondary task (sec) Not significant Number of distinct secondary tasks per traversal Not significant passenger distraction Not significant cognitive distraction Not significant object in vehicle distraction Not significant electronic device distraction NA NA NA NA NA NA vehicle controls distraction Not significant external distraction Not significant eating distraction Not significant other distraction NA NA NA NA NA 25

38 Distraction Type Table 9. Prices Fork Road 95% confidence intervals. Low-crash Mean or Proportion High-crash Mean or Proportion Difference Lower Bound Upper Bound Segment with Higher Distraction Involvement Number of secondary tasks per traversal Low-crash Length of longest duration secondary task (sec) High-crash Number of distinct secondary tasks per traversal Not significant passenger distraction Not significant cognitive distraction Not significant object in vehicle distraction Low-crash electronic device distraction NA NA NA NA NA NA vehicle controls distraction Not significant external distraction Low-crash eating distraction Not significant other distraction NA NA NA NA High-crash In fact, no significant differences were observed in any category between the high- and lowcrash segments on US-460 and only a single secondary task was significant for the VA-7 roadway. This would seem to indicate that the differences observed in the aggregate comparisons are being driven by the VA-28 and Prices Fork roadways. It is interesting to note that in cases in which differences were observed it is difficult to discern generally applicable trends. For example, in Table 4 through Table 9 the direction of difference whether the low- or high-crash segment proportion is larger changes across the different types of secondary tasks within a road segment. However, the direction of difference was consistent across roadway types. For example, the external distraction secondary task was found to be significant on VA-7, VA-28, and Prices Fork, and in each case the proportion was higher on the low-crash segment. The number of observed secondary tasks differs significantly between the low-crash and highcrash segments on the VA-28 and Prices Fork roadways, with the direction of difference toward the low-crash segment. It is interesting to note that on the Prices Fork segments, although drivers were engaged in more secondary tasks on the low-crash segment, the distraction-related glances with the longest duration occurred on the high-risk segment. 26

39 Some of the results are notable for lack of an available comparison or a complete absence of data. On VA-28, US-460, and Prices Fork Rd., the catchall other distraction category was only observed on the high-risk segments. No electronic device distractions were noted on any of the Older Driver segments. SUMMARY OF ANALYSIS As noted in the descriptions of many of the graphs and as can be noted in Table 4 through Table 9 (which present secondary task confidence interval estimates), few significant differences were observed between the high-crash and low-crash segments, with the exception of a handful of secondary tasks. Despite not finding many significant differences, the simple comparison analysis revealed several potentially interesting findings, and these are discussed in the next chapter. 27

40

41 CHAPTER 4. DISCUSSION AND CONCLUSIONS The objectives of this study were twofold. The first was to identify roadway infrastructure differences between high-crash and low-crash intersections and rural road segments that may contribute to the higher crash counts. The second was to determine if drivers exhibit different behavior when traversing a high-crash intersection or rural road segment as compared to when traversing a low-crash intersection or rural road segment. The lack of many significant differences across the road segments for both 100-Car and Older Driver may be due to several different causes. One compelling possibility is that drivers simply may not recognize that they have transitioned from a low-crash to a high-crash road segment. That is, cues from either the infrastructure or driving environment are not sufficient or urgent enough to prompt drivers to significantly alter behavior. Along a similar vein, it is also possible that the low-crash segments that were selected for analysis are too similar to the high-crash segments and do not capture drivers transitioning between different types of road segments. The selection method for low-risk segments was based on the number of crashes reported on various road segments relative to the reported number of crashes on the high-risk segments. It is possible that the relative basis for selection was not sufficient to pair roadways with different characteristics. As always, a lack of significant differences could be due to a lack of statistical power due to using a small number of observations. Unlike controlled experimental comparisons, insufficient power can be a problem with naturalistic data due to the resources required to locate events of interest and complete the video reduction process. The fact that consistent patterns are not apparent due to changing direction of difference may indicate that generally identifying causes for crashes in rural road hotspots may be very dependent on the specific location of interest. If this is the case, this may point to the need for some form of infrastructure-based data collection that is capable of collecting all traversals through a road segment for a specified period of time. This form of collection would allow for customized solutions for each problematic road segment. FUTURE WORK While the outcome of this method has not revealed specific differences between driving behavior and infrastructure in high- and low-crash locations, some possibilities have been identified that could be pursued in the future. Of particular interest is further look into the ability of drivers to identify high-crash intersections and roadway segments. A study to examine the cues that drivers use to identify hotspots, or areas that require different levels of attention or driving techniques could be pursued by examining road segments with similar geometry and traffic patterns but with lower crash rates. What cues are available to drivers that encourage successful and safe traversal in these areas? Are some of the warnings, such as signage, more or less successful than others? Are there other cues that could be provided that are more successful than those that are currently in use? Also of interest is a more detailed evaluation of the different types of secondary tasks the different cohorts engage in. For example, no electronic device secondary tasks were observed 29

42 among the older drivers but this was found to be significant on one of the 100-Car segments. Older drivers did show higher-than-expected rates of secondary task engagement, particularly in regard to external distraction and interacting with an object in the vehicle. As with most safety-related events, problems arise infrequently, when a number of riskincreasing conditions occur at the same time. When aggregating many traversals, the infrequently occurring alignment of conditions into a high-risk situation may be obscured by the larger number of routine situations. For this reason, it may be of value to employ a method that focuses on the sequence of events occurring within the traversals. This approach might investigate, for example, how drivers increase or decrease focus on the driving task when approaching a segment or within a segment. Do numbers of distractions increase or decrease? Within each traversal, what is the rate with which primary task demands, such as braking and negotiating turns, arise? In the course of the current work, the question of whether impairment influences route selection was raised. For example, do drivers who are drowsy or under the influence of alcohol tend to favor rural roads, interstates, or local roads? If so, is this different from the driver s typical route selection? Geospatial tools and naturalistic driving data like those used in this effort would be useful in this additional work. Additional research projects that are of interest include selecting different control segments to compare to the high-crash segments to evaluate the extent to which that control segment selection has on the significance of differences. Other possibilities include increasing the sample size at currently selected locations or selecting a small number of files and watching for transition periods in which the drivers appear to alter behavior and determine if transitional cues can be identified. LIMITATIONS OF THE STUDY As with any research endeavor, there are limitations to the current study. Although naturalistic data collection provides valuable insight into many research questions of interest, it is constrained to the participants in the study and to the locations that the participants traversed. The primary limitation is that this study was restricted to eight roadway segments and a limited number of trips were observed. In addition, the simple comparisons that were used did not explicitly take into account participant-specific differences that may exist. These limitations should be kept in mind when drawing general conclusions from these results. 30

43 APPENDIX A. VIDEO REDUCTION Research Task The purpose of this research effort is to determine if behavioral differences exist when a participant is traversing different roadway types. The following roadway types are under consideration: urban intersections and rural roadway segments. Reduction Task Each epoch included in this reduction will consist of a single traversal across a predetermined roadway segment. The reduction questions are grouped as follows: 1. questions regarding event spot-checking and video quality; 2. questions that use entry to roadway segment (i.e., the first sync of reduction epoch) for timing information; 3. questions regarding the participant (behavior, impairment, driving style) and the driving environment (lighting, weather, visual obstructions); 4. questions relevant only to intersection traversals. In order to capture participant actions and behavior, multiple opportunities are provided for recording values for some of the variables. In the event that multiple values are recorded for a variable there is no need to attempt to rank the importance of the multiple values. For example, if a participant is adjusting radio controls and eating a sandwich there is no need to attempt to decide which is having a larger impact on driving performance. The remainder of this manual lists the questions, responses, and answer guides for this question reduction. It is strongly recommended that you familiarize yourself with the questions and responses prior to beginning reduction. Reduction Questions 1. Spot checking Has this event been spot checked? Yes No 2. Video quality forward view, face view, over the shoulder view, rearward view rear windshield, rearward view passenger window The purpose of the video quality questions is to determine if reduction should continue on the current epoch and to provide an indication of video quality for a given file. If the forward view, face view, or over-the-shoulder view video is missing or of unacceptable quality (including severe obstruction or misalignment) skip the current epoch and proceed to the next epoch. These views, at minimum, should be marked as present and usable. 31

44 If one or both of the rearward views are less than present and usable but the other views are acceptable, complete the event as best as possible. Present and usable the video view is present with good video quality and free of obstructions and misalignment. For the face view, this means that both eyes and mouth are visible for at least 75% of the event. Present with poor video quality the video view is present but with degradation of video quality. Present but obscured the video view is present but obscured. Obstructions might include the sun visor, a hat, or lighting conditions. Present but misaligned the video view is present but the camera is not properly aligned. Not present the video view is not present. 3. Lane of travel what was the participant lane of travel on entry to roadway segment? 1 thru 8 o The reference point for the lane of travel variable is the shoulder to the right of the vehicle. The lanes are number from right to left, [roadway segment only] Other vehicles how many other vehicles were present around the subject during roadway traversal? Only consider other vehicles that are close enough to restrict the participant s ability to maneuver. Count each vehicle a single time. Only answer this question for roadway segments; separate questions exist for other vehicles during intersection traversals. 5. Infrastructure Which of the following roadway elements seemed to impact the participant s ability to navigate the roadway segment? None Roadway alignment maneuver is difficult for the driver due to the geometry of the roadway or intersection o ex. Narrow roadway, sharp turn, short merge lane Roadway sight distance the driver is unable to see an adequate distance ahead Traffic control device lack of, positioning of, or poor visibility of traffic control devices Roadway delineation poor visibility or positioning of roadway or lane borders o ex. Faded paint on lane lines Weather, visibility roadway is not designed to accommodate certain weather conditions and thus weather conditions influence driver capabilities o Ex. Lane markings are difficult to distinguish on wet pavement Other o This category would include infrastructure elements such as construction barrels or cones 32